As AI continues to demonstrate its utility in an astonishing array of medical imaging applications, human leadership can partner with AI to build a better system of lifelong learning, according to Elizabeth S. Burnside, MD, MPH.
Burnside
"It's important to intentionally preserve the core manifestations of human intelligence, especially those at the heart of patient care—including judgment, transparency and communication," said Dr. Burnside, a professor and associate dean at the University of Wisconsin School of Medicine and Public Health and co-executive director of the university's Institute for Clinical and Translational Research.
In her plenary lecture, Dr. Burnside showed a series of examples demonstrating key points in the "imaging chain" where AI can offer benefits to each member of the medical imaging team—and illustrated areas where AI may still be problematic.
With a simplistic but illuminating exercise, she tasked an AI program with producing images of Marie Curie, Wilhelm Roentgen and Godfrey Hounsfield in Chicago Cubs uniforms at Wrigley Field. Within the resulting images, the iconic faces, uniforms and stadium were mostly recognizable, but the program had also inserted incongruous structures, placed "fans" on precarious ledges and generated a Hounsfield with three legs.
"Discriminative AI models are primarily used to classify existing data into predetermined outcomes of interest," Dr. Burnside explained. "Generative models, on the other hand, use algorithms to craft high-quality—we hope—content, including text and images based on the data on which they were trained."
In a radiology setting, examples of discriminative AI tasks could be analyzing a chest X-ray and determining whether pneumonia is related to COVID-19 infection, identifying cancer on a mammogram or finding a bleed on a neuroimaging study. A generative model could use images to create a report, improve performance on image segmentation, simulate disease progression in a body system or create summaries for patients in lay language.
Current Attitudes Toward AI in Radiology
Dr. Burnside presented results from several membership surveys, conducted by both U.S. and European societies, giving the audience a snapshot of members' views on implementing AI in the radiology arena. She also unveiled findings from a yet-unpublished survey that concluded last month, among the Society of Chairs in Academic Radiology Departments (SCARD). The results illustrate respondents' level of optimism about different forms of AI, their interest in its potential applications—such as alleviating burnout—and the cost of its implementation versus the value.
Ninety-three percent of respondents said they were optimistic about AI in general, while 86% were optimistic about generative AI. Asked to rank the importance of various AI applications, the chairs unanimously identified quality and efficiency as "very important" or "extremely important," and a similar number found it "extremely important" to address burnout. Lower in the rankings were salaries, cost and education, but more than half of respondents said that equity was important or extremely important.
"I laid these factors out on a spectrum of 'tame problem' versus 'wicked problem," Dr. Burnside said, placing quality/efficiency at the tame end and equity at the wicked end. "So kudos to chairs who are making sure our AI algorithms consider equity."
She then mapped out a representation of applications according to their place on the imaging chain—pre-order, order, pre-imaging, acquisition, post-processing, interpretation workflow, care coordination and downstream workflows. A clear pattern emerged, with respondents prioritizing post-processing, interpretation and acquisition, respectively.
"We must pay attention to all aspects of the imaging chain, as well as to existing governance structures," Dr. Burnside said. "We'll need a holistic approach to develop a framework designed to maintain a culture of stakeholder empowerment, collaboration and continuous learning."
Access the plenary session, "Leading Through Technology: Valuing Artificial and Human Intelligence," (M4-PL02), which includes additional results from the new survey, on demand at Meeting.RSNA.org.
